Combination of validity aggregation and multi-scale feature for person re-identification

Person re-identification (re-ID) gradually has attracted attention of the industry and academe. It has a great application future in computer vision with the development of deep learning. The main challenges in the process of studying re-ID come from different camera angles, occlusion, and person’s posture changes. How to extract a powerful person descriptor is a fundamental problem in re-ID task, which is still an open topic today. In this study, we propose a Validity aggregation and multi-scale feature extraction network (VMSFEN), based on global and local features to convey more interesting information. In order to tackle local feature misalignment and feature pairs contribution, a novel strategy that combines cross-alignment and validity aggregation strategy is embedded into our model. Cross-alignment aims to obtain the same semantic features according to max semantic features. Validity aggregation provides appropriate weight to each matched local feature pair. Finally, we integrate the learned feature pairs with calculated weight, and introduce it into the triplet loss function. The approach of this work achieves 96.3% Rank-1 and 88.6% mAP on Market1501, 89.9% Rank-1 and 80.3% mAP on the DukeMTMC-reID, 80.1% Rank-1 and 76.3% mAP on the CUHK03. These findings prove that VMSFEN is an efficient network in re-ID study.